In an earlier article I argued that art education has a legitimate claim to being one of the most important sites for developing critical thinking in students — but only if we’re deliberate about building it in. This is an account of one attempt to do exactly that, using AI in ways that might surprise people who assume the conversation about AI in art class begins and ends with image generation.
In 2024 at HKUGA College in Hong Kong, I introduced an AI-assisted learning tool into the junior form visual art curriculum with two specific goals. The first was to help students develop their ability to use visual evidence to support analytical arguments about artworks. The second was to support students who were emerging English writers — a significant portion of my S1 to S3 classes — without turning the tool into a grammar corrector.
The platform I used was POE, which allowed me to build a custom assignment environment and program a teaching assistant I called Yue. The interface was straightforward: students were presented with artworks and an analytical question, wrote their initial response, and then clicked a button to activate Yue. Yue would analyze their response, return specific suggestions for how to strengthen their use of visual evidence, and provide an estimated score. Students then wrote a revised response based on those suggestions. Importantly, Yue did not address grammar or sentence structure — the feedback was focused entirely on the quality and specificity of the visual evidence being used to support the argument.
The resource buttons embedded in the platform — Vocab Help, Elements of Art, Principles of Design, Theme Help, Glossary — meant students had immediate access to the analytical vocabulary they needed without leaving the assignment environment.
What the screenshots below show is the feedback loop in action. A student’s original response to a question comparing the moods projected by comic book covers from different decades describes what they see — characters, settings, general impressions — but relies almost entirely on personal feeling rather than visual evidence.
Yue returns a score of 15/18 and two specific suggestions: to discuss how stylistic elements — line, shape, color, texture, space — actually create mood rather than simply reporting what the mood feels like, and to use comparative language to analyze how different artists achieve different effects.
The revised response shows a meaningful shift. The student begins connecting specific visual observations — “the background colour purple and yellow make a strong contrast,” “the lines are thin and detailed” — to their analysis of mood. It isn’t a perfect response, but the student is now doing something categorically different from what they were doing before.
The value of this feedback loop is partly about quality and partly about timing. Research on AI-assisted writing feedback consistently identifies immediacy as a critical factor — students who receive feedback while still engaged with a task revise more substantively than those who wait a week for teacher comments by which point the original thinking has evaporated. A teacher working with thirty students simply cannot replicate that cycle at scale.
Over the course of the semester, scores on revised responses increased by approximately 20% compared to original responses. I want to be careful about that number. Students typically completed only two or three of these assignments per semester, which is not enough to draw firm conclusions about the long-term internalization of these skills. What it suggests, cautiously, is that the immediate feedback loop was useful in the moment. Whether those habits of mind transferred and accumulated over the three-year junior form period is harder to measure — though anecdotal evidence from later classes suggested that students were more consistently grounding their arguments in visual observation than comparable groups had been previously. That could be the feedback loop. It could also be their growing familiarity with the format, the vocabulary resources, or the supporting videos I produced for each project. Probably all of the above.
The next iteration of this tool would move beyond the current model toward something more genuinely dialogic — a platform where the AI guides students through a structured observation process, helps them develop those observations into an analytical response, evaluates that response, and then has them write a final version independently. The student would begin with a short multiple choice observation activity about an artwork; the AI would then differentiate its guidance based on the student’s demonstrated ability level before the analytical writing begins. That’s a more sophisticated feedback loop than POE can currently support, but the architecture is clear. The broader point is this: the conversation about AI in art education needs to move past image generation. The more interesting question is whether AI can help students learn to look more carefully, argue more precisely, and think more rigorously about what they see. At least in one classroom in Hong Kong, the early evidence suggests it can